| Summary: | 碩士 === 大同大學 === 資訊工程學系(所) === 104 === This research is a continuing study of the language training system for the hearing impaired developed in our laboratory. The purpose is to investigate the rating mechanism of the training system so that the evaluation results will be more accurate.
We recorded a corpus database uttered by six people, including three males and three females. The speech features used in this research including MFCC, LPC, LPCC and PLP. The dynamic time wrapping algorithm is used to calculate the similarity score between reference speech and testing speech. The similarity score was a weighted sum of the DTW distance score and DTW slope score.
Experiments were composed of three parts. In the first two parts, we compared the relation of the rating factors of the DTW distance score and DTW slope score. In the last part, we compared the similarity score by varying the weights for the two rating factors. The results show that in the agree/disagree similarity measure, the average accuracy is 68.5%. In one level adjacent similarity measure, the average accuracy is 63%. The similarity rate with MFCC feature is better than those with other features. With MFCC feature, the accuracy in agree/disagree similarity measure can achieve 70%.
|
|---|
